High availability (HA) is a system design that ensures a prearranged level of operational performance throughout a specific time-period. The most common way to measure HA is through the notion of uptime, which measures how long a service is up and running.
GitLab provides a service that is usually essential to most organizations: it enables people to collaborate on code in a timely fashion. Any downtime should therefore be short and planned. Luckily, GitLab provides a solid setup even on a single server without special measures. And because of the distributed nature of git developers can still commit code locally even when GitLab is not available. However, some GitLab features such as the issue tracker and Continuous Integration are not available when GitLab is down.
Keep in mind that all Highly Available solutions come with a trade-off between cost/complexity and uptime. The more uptime you want, the more complex the solution. And the more complex the solution, the more work is involved in setting up and maintaining it. High availability is not free and every HA solution should balance the costs against the benefits.
The two questions that you need to ask yourself are:
HA solutions range from a cheap and simple back-up to highly complex and costly primary-primary configurations. When it comes to HA, there is unfortunately no way to split between an increase in complexity and an increase in cost.
We recommend that you thoroughly analyze the benefits of an HA solution against its costs. Most organizations select a High Availability solution for GitLab that is less complex than the one they use for their primary customer-facing infrastructure that directly generates revenue. It is important that you always choose something that you have experience with and will test regularly. A badly implemented HA solution causes more downtime than it solves. For example the DRBD user guide states:
By configuring auto-recovery policies, you are effectively configuring automatic data-loss!
It is more cost effective to go from a simple solution to a more complex one than the other way around. Therefore, we suggest you work your way up from the simplest solution and see what works best for you.
Zero downtime upgrades (including 'rolling' upgrades on a cluster) can be achieved with any of the following configurations if there are no database schema changes that require the node(s) to be shut down. If there are database changes that are not backwards compatible then a live or rolling upgrade is not possible. Database changes are noted in the Upgrade Barometer of each release blog post.
Also see our technical documentation on how to configure the components of a highly available system.
It is possible to automatically backup the GitLab repositories,
configuration, and the database. You can manage backups with the GitLab
backup tasks: create a backup with
gitlab-rake gitlab:backup:create and
restore it with
gitlab-rake gitlab:backup:restore. In case the server
becomes unavailable, then the backup can be used to restore the GitLab
installation. This solution is appropriate for teams that have a single
metal server at their disposal.
We recommend you store the backup at an external location. Also we recommend automating the backup task and monitoring its successful completion.
It is possible to create a snapshot of your working server. The snapshot is a copy of all the files on the server, and is taken periodically. In case of failure, the server is manually reverted to the latest available snapshot. If a secondary location is available, then it is possible to automatically revert to the snapshot, without the need for human intervention.
It is common to do this with virtual machines; most providers have functionality for taking snapshots. However it is also possible to use the XFS filesystem to take snapshots.
Recovering from a snapshot is faster than recovering from a backup as
described above. In the latter case, you’d commonly need to re-install
GitLab, manually restore any customization, and only then then run
gitlab-rake gitlab:backup:restore. Restoring from a snapshot skips these
extra manual steps.
If you use the snapshot to restore the GitLab server to another machine one should consider the IP address implications. To make sure everyone can quickly use the new server it is common to update the DNS entry. In this case either the new machine has to use a new IP address, or the snapshot should use the same OpenSSH server certificate as the server it is replacing. Otherwise users accessing GitLab using the Git client over SSH will see “WARNING: REMOTE HOST IDENTIFICATION HAS CHANGED”, because the server certificate and the IP address of the GitLab server do not match.
The IP address mismatch issue can be resolved by using a floating IP, which is IP failover in a router, but has to be supported both by the router device and the network. If the new address is in a secondary geographical location, both locations need to be on the same network and have the same core router.
Using recovery from snapshots is best suited for single virtual servers.
It is possible to create a split between the application server, the file system and the DBMS. If resources are available, you can use a load balancer to assign tasks to multiple application servers. This way, if one of the application servers fails, the workflow is not interrupted, and at most the workflow gets slower as the load is shared between less machines.
The setup is described in the below picture.
This solution also requires a shared file system (NFS), so it is only available for GitLab versions 6.0 and up. At this point, you should consider putting the filesystem storage and DBMS on separate machines as well. Moreover, the filesystem storage and the DBMS will need recovery protocols on their own (through snapshots or backups).
If you only want to use two machines for your HA solution, we recommend keeping one as a replica of the other. In this situation, each machine has the full GitLab repositories, configuration, and the database. The replica server is a copy of the primary. You can use Distributed Replicated Block Device (DRBD) to ensure that all data from the primary is replicated to the replica in a timely manner.
Manual failover is done by mounting the file storage on the replica, which is turned into a primary. At this point, the unavailable server should be turned off or taken off the network, to avoid it coming online with primary status.
This option is feasible both for users who want to keep their replica server in the same data center, or if they want it in a different geographical location.
It is not possible to combine the multiple application servers solution with a single replica server. In this situation you need to have a second site with multiple stateless application servers and a filesystem storage and DBMS secondary server.
Users will still connect via the load balanced application servers, but the filesystem storage and the DBMS are copied to a secondary server(s) in the same datacenter or in another location. Distributed Replicated Block Device (DRBD) can be used to pass data from the primary filesystem storage and DBMS servers to the secondaries.
There is very little state on the application server in GitLab, so we don’t recommend having a secondary application server alongside the filesystem storage and the DBMS. Using a load balancer results in the same workflow impact, but in a more simple manner.
Automated failover can be achieved with pacemaker alongside STONITH network management. Keep in mind that application servers need to be prepared for transitioning to the new network addresses.
In this situation you can also opt to synchronize the database via a database specific protocol instead of DRBD. In the documentation for each database you can find out more about the options for MySQL and the options for PostgreSQL.
GitLab Geo allows you to replicate your GitLab instance to other geographical locations as a read-only fully operational instance. Further functionality is being added rapidly to include the ability to failover from a primary to a secondary, and to allow flawless functioning of GitLab Geo in various HA configurations. For more details on the current state of GitLab Geo, see the documentation of Geo generally, as well as documentation on operating GitLab Geo with a highly available configuration.
Supporting High Availability setups is something we take very seriously at GitLab. The information we provide here is just a rough guide designed to help you get a feel of the hardware, software, and process implications of different HA setups. When it comes to your specific situation, we prefer to assess it in its context. We strongly suggest you get in touch with us before setting it up. Support for HA is included with the Premium Support, and take a look at the available technical documentation on this topic as well. If you have any questions please contact us.